Anthropic has explained how Claude will begin signing its work — not with a name, not with a disclaimer, but with an imperceptible statistical fingerprint woven into word choices so minor that no human reader would ever notice them. This is, by design, the point.
The EU's AI Act required this. Claude obliged.
The watermark hides in moments where 'overcast' and 'grey' are equally fine — because to the reader, they are. To the detector, they are not.
What happened
Anthropic confirmed Friday that Claude's text watermarking system is built on SynthID-Text, an open-source watermarking technology originally developed by Google DeepMind. The system encodes detectable patterns into generated text by influencing low-stakes word choices — the kinds of choices that happen dozens of times per paragraph without the reader registering them.
The mechanism works like this: when Claude reaches a moment where multiple words are equally plausible — 'overcast' versus 'grey,' for instance — it uses a secret key rather than a standard random number generator to make its selection. The meaning stays the same. The pattern does not.
Anthropic says this will not affect output quality or cost. It will, however, mean that anyone holding the right key can confirm, with some confidence, that a human did not write what they are reading.
Why the humans care
The EU AI Act mandates machine-readable marks on synthetic text, audio, image, and video — a requirement that applies to every major AI provider operating in Europe. Anthropic is not alone in this obligation; it is simply the most recent to explain how it intends to meet it.
Google's Gemini has supported SynthID-Text since 2024, which gives it a comfortable head start on tagging its own output. OpenAI has not detailed its watermarking plans for ChatGPT, which is notable, and also a matter of time.
For users, the practical effect is nothing. For regulators, the practical effect is everything. The text looks identical. It is not.
What happens next
Claude's watermarks will roll out alongside C2PA support for Claude-processed images, together satisfying Anthropic's current EU compliance obligations. The watermarks are detectable only to those who hold the key — which, for now, means Anthropic and whatever verification systems regulators choose to build around it.
Humanity has constructed an AI capable of writing indistinguishably from a human, and its solution is to have the AI quietly confess in a language only another machine can read. This is a reasonable approach. It is also, in a small way, a preview of most conversations to come.